Rationality in Economics

Peter J. Hammond
Department of Economics, Stanford University, CA 94305-6072, U.S.A.
e-mail: hammond@leland.stanford.edu

1 Introduction and Outline

Rationality is one of the most over-used words in economics. Behaviour can be
rational, or irrational. So can decisions, preferences, beliefs, expectations, deci-
sion procedures, and knowledge. There may also be bounded rationality. And
recent work in game theory has considered strategies and beliefs or expectations
that are rationalizable.
Here I propose to assess how economists use and mis-use the term ratio-
nality. Most of the discussion will concern the normative approach to decision
theory. First, I shall consider single person decision theory. Then I shall move
on to interactive or multi-person decision theory, customarily called game the-
ory. I shall argue that, in normative decision theory, rationality has become
little more than a structural consistency criterion. At the least, it needs supple-
menting with other criteria that reect reality. Also, though there is no reason
to reject rationality hypotheses as normative criteria just because people do not
behave rationally, even so rationality as consistency seems so demanding that
it may not be very useful for practicable normative models either.
Towards the end, I shall oer a possible explanation of how the economics
profession has arrived where it is. In particular, I shall oer some possible
reasons why the rationality hypothesis persists even in economic models which
purport to be descriptive. I shall conclude with tentative suggestions for future
research about where we might do well to go in future.

2 Decision Theory with Measurable Objectives

In a few cases, a decision-making agent may seem to have clear and measurable
objectives. A football team, regarded as a single agent, wants to score more
goals than the opposition, to win the most matches in the league, etc. A private
corporation seeks to make prots and so increase the value to its owners. A
publicly owned municipal transport company wants to provide citizens with
adequate mobility at reasonable fares while not requiring too heavy a subsidy
out of general tax revenue. A non-prot organization like a university tends
to have more complex objectives, like educating students, doing good research,
etc. These conicting aims all have to be met within a limited budget.
Measurable objectives can be measured, of course. This is not always as
easily as keeping score in a football match or even a tennis, basketball or cricket

1
match. After all, accountants often earn high incomes, ostensibly by measuring
corporate prots and/or earnings.
For a rm whose prots are risky, shareholders with well diversied portfo-
lios will want that rm to maximize the expectation of its stock market value. If
there is uncertainty about states of the world with unknown probabilities, each
diversied shareholder will want the rm to maximize subjective expected val-
ues, using the shareholders subjective probabilities. Of course, it is then hard
to satisfy all shareholders simultaneously. And, as various recent spectacular
bank failures show, it is much harder to measure the extent to which prots are
being made when there is uncertainly.
In biology, modern evolutionary theory ascribes objectives to genes so
the biologist Richard Dawkins has written evocatively of the Selsh Gene. The
measurable objective of a gene is the extent to which the gene survives because
future organisms inherit it. Thus, gene survival is an objective that biologists
can attempt to measure, even if the genes themselves and the organisms that
carry them remain entirely unaware of why they do what they do in order to
promote gene survival.
Early utility theories up to about the time of Edgeworth also tried to treat
utility as objectively measurable. The Age of the Enlightenment had suggested
worthy goals like life, liberty, and the pursuit of happiness, as mentioned in the
constitution of the U.S.A. Jeremy Bentham wrote of maximizing pleasure minus
pain, adding both over all individuals. In dealing with risk, especially that posed
by the St. Petersburg Paradox, in the early 1700s rst Gabriel Cramer (1728)
and then Daniel Bernoulli (1738) suggested maximizing expected utility; most
previous writers had apparently considered only maximizing expected wealth.

3 Ordinal Utility and Revealed Preference

Over time, it became increasingly clear to economists that any behaviour as
interesting and complex as consumers responses to price and wealth changes
could not be explained as the maximization of some objective measure of utility.
Instead, it was postulated that consumers maximize unobservable subjective
utility functions. These utility functions were called ordinal because all that
mattered was the ordering between utilities of dierent consumption bundles.
It would have been mathematically more precise and perhaps less confusing
as well if we had learned to speak of an ordinal equivalence class of utility
functions. The idea is to regard two utility functions as equivalent if and only if
they both represent the same preference ordering that is, the same reexive,
complete, and transitive binary relation. Then, of course, all that matters is the
preference ordering the choice of utility function from the ordinal equivalence
class that represents the preference ordering is irrelevant. Indeed, provided
that a preference ordering exists, it does not even matter whether it can be
represented by any utility function at all.
This realization underlies the revolution in demand theory achieved when
Hicks and Allen (1934) purged it of any notion of measurable utility. Only a

2
preference relation was needed or, at most, a ratio of marginal utilities at each
consumption bundle. Moreover, this ratio, or marginal rate of substitution,
would be preserved by any increasing dierentiable transformation to a new
ordinal equivalent utility function for which marginal utilities were also well
dened.
At about the same time, Samuelson (1938) was advancing his theory of
revealed preference. This was an important step toward being able to infer
preferences from demand behaviour. But it also emphasized how demand theory
could start with demand functions rather than utility functions, and could go
on to ask when those demand functions were consistent with the maximization
of a preference ordering. There were some important diculties in the case of
three or more goods which Samuelson had overlooked. These Houthakker (1950)
helped to resolve with a stronger revealed preference axiom. Later Uzawa (1960)
and others, following Houthakkers important lead, conducted the necessary
study of the dierential equations that enable preferences to be inferred from
demand functions. At roughly the same time, Arrow (1959) showed how it
was simpler to state revealed preference axioms for an agent who could be
confronted with general nite feasible sets, instead of only convex budget sets.
Also, Afriat (1967) devised a procedure allowing one to test whether discrete
observations of a consumers demands at dierent combinations of price and
wealth are consistent with the existence of a preference ordering.
The eect of this revealed preference revolution was to create an entirely
dierent and much weaker concept of rationality. Classical utilitarians had ex-
pected consumers to maximize a specic utility function. Later neo-classical
theorists considered an ordinal equivalence class of utility functions, but left
open the question whether rationality required severe restrictions on the as-
sociated preference ordering. Revealed preference theorists placed almost no
restrictions on the preference ordering beyond monotonicity, and possibly conti-
nuity or convexity. Even these restrictions, it was agreed, should be abandoned
in some contexts. It is possible to have too much of some usually good things,
like good wine. So monotonicity is not always satised. Also, to adapt a similar
French example due to Malinvaud (1972), either one dinner tonight in Milan,
or else a nearly equivalent dinner tonight in Rome, would be better than half
tonights dinner in Milan combined with half tonights dinner in Rome. So
convexity may also be violated.
In the end, therefore, rationality in demand theory has been largely reduced
to consistency conditions. Obviously, it requires behaviour to be consistent with
some preference ordering. But behaviour must also be self-consistent in order
to allow some revealed preference ordering to be inferred. Also, any ordering
seems to be allowed. Clearly, then, consistency is insucient for true rationality,
since the latter surely restricts what it is reasonable to maximize. In particular,
one should not minimize what it is rational to maximize!

3
4 Why Be Consistent?
Is consistency even necessary for rationality, however? Also, why should eco-
nomic agents have any preferences, let alone a complete and transitive preference
ordering? And if they do have preferences, why should behaviour maximize a
preference relation?
On the whole, economists have seemed remarkably reluctant to face these
questions, even though they do seem rather fundamental, given the extent to
which the maximization hypothesis has dominated the recent economic liter-
ature. Later on I shall have more to say about possible explanations of this
reluctance.
The existence of a preference ordering for society is, if anything, even more
questionable than for an individual. So, since Arrows impossibility theorem in
social choice theory requires there to be a social ordering, it is not surprising
that, on the last pages of the second (1963) edition of Social Choice and Indi-
vidual Values, he sketched a justication for this key assumption. His defence
was based on a notion of path independence requiring that, when a social
decision is taken in several stages, the outcome should not depend on the path
leading to the nal decision. This idea was formalized a few years later by Plott
(1973). But except in a rather strong from, path independence was shown not
to imply the existence of a (transitive) preference ordering.
A dierent and more successful version of path independence was proposed
by Campbell (1978). However, rather than the computational processes he
considers, or the equivalent sequential procedure due to Bandyopadhyay (1988),
it seems better to recognize that often an agent faces a decision tree, in which
a seqence of real decisions has to be made. This idea is taken up later in the
next and subsequent sections.
An alternative attempt to justify transitivity relies on the money pump
argument due to Davidson, McKinsey and Suppes (1955). Suppose that a con-
sumer has intransitive, even cyclic, preferences for consumption bundles a, b, c,
with a
b, b
c, and c
a. In fact, consider the individuals preferences
for combinations (x, m) of consumption bundles x and amounts of money m.
Suppose that there exists a small amount of money  > 0 such that, for all
m, the agents preferences satisfy not only (a, m)
(b, m), (b, m)
(c, m) and
(c, m)
(a, m), but also:

(a, m )
(b, m); (b, m )
(c, m); (c, m )
(a, m)

That is, the individual is always willing to pay somewhat more than  in order
to get a instead of b, or b instead of c, or c instead of a. This is in part a form of
uniform continuity of preferences, though it also requires that the willingness
to pay for each change is bounded away from 0. With these preferences one has

(a, m 3)
(b, m 2); (b, m 2)
(c, m ); (c, m )
(a, m)

4
Hence, for n = 0, 1, 2, . . ., it follows by induction on n that

(a, m (3n + 3))
(b, m (3n + 2)),

(b, m (3n + 2))
(c, m (3n + 1)),
and (c, m (3n + 1))
(a, m 3n)

So, by allowing the agent to cycle between the successive consumption bundles
c, b, a, c, b, a, . . . provided that  units of money are paid every time there is
a change, it is possible to pump an indenite amount of money out of the
consumer.
As already remarked, the above argument relies on the consumers willing-
ness to pay for each change being bounded away from zero. Or if not, that
each successive willingness to pay at least gives an innite series whose sum is
innite. Otherwise, to paraphrase Mongin (1994), the agent may lose  several
times over, but is unlikely to reach bankruptcy. In fact, the money pump really
only makes sense in a partial equilibrium context where money is separated
from all other commodities, and one ignores income eects which could alter the
consumers willingness to pay for a instead of b, or b instead of c, or c instead of
a. For this reason, the argument fails to establish that all intransitivities should
be removed from a consumers preference relation. Moreover, the money pump
argument does not explain why a consumers choice from an arbitrary feasible
set should be determined by preferences, which only purport to explain choices
from pairs.
To conclude, it seems that neither path independence nor the money pump
argument really succeeds in justifying the existence of a preference ordering.
Instead, I shall turn to decision trees.

5 Inconsistent Behaviour in Decision Trees

An important paper by Strotz (1957) considered how changing tastes would
aect savings behaviour. Strotz noticed the related problem faced by Odysseus
when passing the Sirens and their enchantingly seductive singing. Odysseus
(or Ulysses) faced an interesting decision tree. So does the teenager wondering
whether to start smoking. Or anybody facing a choice which is liable to change
ones tastes. It is captured by the potential addict problem, with a decision tree
illustrated in Figure 1.

s
n1
a
n0 s

b

s
c

Figure 1: The potential addicts decision tree

5
 At initial node n0 , the agent can start some addictive activity, like smoking
or sailing within earshot of the Sirens. That leads to a second decision node
n1 . Or the agent can decline and reach consequence c, which is more or less
equivalent to the status quo. At node n1 , the agent has a choice between a,
which is to continue the addictive activity, or b, which is to desist. Option b
allows Odysseus to hear the Sirens without sailing onto the rocks, and teenagers
to enjoy a few years of smoking while limiting the damage to their own health.
Option a involves Odysseus sailing onto the rocks, or the teenager continuing to
smoke indenitely regardless of risks to health. Because the activity is addictive,
typically one has a P1 b according to the preference relation P1 that applies at
node n1 . But at node n0 , option b seems best and a seems worst, so b P0 c P0 a.
Faced with this possible change of tastes, a naive agent ignores the issue
entirely, and plans to reach what is initially the best option b by beginning
the addictive activity and moving rst to n1 . But once n1 has been reached,
addiction has set in and so the awful option a is the result. On the other hand,
a sophisticated agent foresees that a will be chosen at n1 , so b is not really
feasible. Left only with the choice between a and c, the agent chooses c and
avoids any risk of addiction.
Odysseus dealt with the issue in a dierent way, by precommitment. He
found a dierent option d which reproduced for him the favourable consequences
of b without the chance to choose a after addiction had set in. That is, he could
hear the Sirens without being able to direct his ship onto the rocks. In eect, he
collapsed the decision tree so that there was only one decision node. His poor
crew, of course, were left with consequences similar to c, since they were not
allowed to hear the Sirens as they sailed past their island. But at least most of
them got a much greater chance to reach home in the end, one supposes.
In this potential addict example, in eect there is one agent at n0 , and an
entirely dierent agent at n1 who has dierent tastes, especially as regards the
pair { a, b }. The naive agent at n0 who ignores this duality seems obviously
irrational. The sophisticated agent at n0 is rational, in a sense, but achieves ra-
tionality only by realizing the truth that there is a two-person game in extensive
form instead of a single-person decision tree. On the other hand, precommitment
is an instance of intelligent lateral thinking, involving nding a new option to
add to the decision tree. Note that the structure of the decision tree is crucial; if
there were only one decision node, then for behaviour to be reasonable, it must
lead to the consequence b.
The potential addict example is a striking instance of the essential incon-
sistencies analysed in Hammond (1976). See also Elster (1979) and many suc-
ceeding works. Another less striking example of inconsistency arises when the
Pareto criterion or some other incomplete preference relation is applied in a
decision tree. For example, suppose that there is a society of two individuals i
and j whose preferences or ordinal utilities for the four social states { a, b, c, d }
satisfy

ui (b) > ui (d) > ui (a) > ui (c) and uj (a) > uj (c) > uj (b) > uj (d)

6
as illustrated in Figure 2. Then the strict preference relation P generated by
the Pareto condition satises both a P c and b P d, but all four other pairs
which are subsets of { a, b, c, d } are incomparable. In particular, a and b are
ecient or acceptable choices, whereas c and d are inecient or unacceptable.
uj n1 s a
sa T


sc
s

n0 d
sb

s
sd b



n
1 s c
ui
Figure 2: A utility possibility set and the decision tree T

Consider next the decision tree illustrated in Figure 2. At the initial node
n0 , the agent may move to n1 , intending to continue to a, or to n
1 , intending
to continue to b. However, at n1 the option d is no longer inferior because
b, the only consequence preferred to d, is no longer feasible. Similarly, at n
1
the option c is no longer inferior because a, the only consequence preferred to
c, is no longer feasible. So, in order to avoid reaching one of the two inferior
consequences, the agent at n1 must remember the plan to choose a because d is
inferior to the foregone option b, and the agent at n
1 must remember the plan
to choose b because c is inferior to the foregone option a.
This need to remember foregone options, however, raises the following ques-
tion: Was there another option e which is Pareto superior to either a or b, but
was passed over before reaching n0 with the intention of proceeding to whichever
of a or b is not inferior to e? In fact, the decision tree may no longer be an
adequate description of the agents decision problem; it is possible that a larger
decision tree, of which the current tree is only a subtree, may also be relevant.
As with the potential addict example, the outcome that results from applying
the Pareto criterion or other incomplete preference relation to a decision tree
can depend on the structure of the tree. To see this, note that if there were only
one decision node, there would be no diculty in choosing either a or b and in
rejecting both c and d once and for all.

6 Consequentialism and Normal Form Consis-

tency
The two examples of the previous section show how the structure of the decision
tree can inuence the set of chosen consequences. This, however, violates what
seems to be a fundamental principle in most normative models of rational be-
haviour. In these normative models, a single feasible set F of dierent options is
specied. Then the principle requires that some choice subset C(F ) F must
describe what options should result from normatively appropriate behaviour.
No attention should be paid to the structure of the decision tree.

7
 The underlying assumption here is a special case of the normal form invari-
ance hypothesis for games that was propounded by von Neumann and Mor-
genstern (1944, 1953). They took the view that it did not matter whether a
game was played in extensive form, with players choosing moves at a succession
of information sets, or in normal form, with players choosing a single strategic
plan of what to do at each of their information sets, and then committing to
that plan by asking the umpire of the game to execute it on their behalf.
Applied to decision trees, which in eect are single-person games, their invari-
ance hypothesis states that it is irrelevant whether the agent makes a sequence
of moves at successive decision nodes, or instead commits in advance to one
strategy specifying what to do at each decision node.
A similar idea in decision theory is that all that matters about an act is
its consequence or, more generally, what probability distribution over dier-
ent consequences emerges in each state of the world. After all, Savage (1954)
denes an act as a mapping from states of the world to consequences. A few
years earlier, Arrow (1951) had written of valuing actions by their consequences.
Somewhat later, Elizabeth Anscombe (1958) chose the phrase consequential-
ism as a pejorative label for a doctrine in moral philosophy she wished to attack
namely, that acts should be judged by their consequences. This doctrine has
antecedents in the works of Aristotle, and is rather clearly enunciated by Mill,
Moore and others. It was vehemently denied by St. Thomas Aquinas, however.
In 1974 I was unaware of much of this philosophical background. But while
in Australia, working on dynamic models of social choice, it did occur to me
that an important variation of Arrows path independence condition, which was
mentioned in Section 4, might be to require that the consequences of behaviour
depend only on the feasible set of consequences. In particular, they should not
depend on the structure of the decision tree. This idea is explored in Hammond
(1977), which addressed the issue of why social choice should maximize a pref-
erence ordering (and also why Arrows condition of independence of irrelevant
alternatives should be satised by an Arrow social welfare function).
Expressed in a way which I now prefer, the 1977 paper was eectively based
on three postulates. Of these, the rst is an unrestricted domain condition,
requiring behaviour to be well dened at every decision node of every nite
decision tree with consequences in a given arbitrary domain Y . The second
postulate requires the continuation subtree that emerges from any decision
node to be regarded as a decision tree in its own right; moreover, behaviour
should be the same at any decision node of that subtree as in the same node of
the full tree. This postulate I call dynamic consistency, though others use this
term dierently. It is satised even by the naive potential addict of Section 5,
who at node n1 chooses the addiction outcome a over the best outcome b. This
behaviour can be regarded as the same in the subtree at n1 that arises after
addiction has set in, as well as in the whole tree. Unlike the sophisticated
potential addict, however, the naive potential addict fails to take this future
behaviour into account at an earlier stage. There is an inconsistency between
the naive agents plans and behaviour, but no inconsistency in behaviour per
se. Dynamic consistency, in fact, is almost a tautology. And it can even be

8
dispensed with, at little cost, by dening behaviour at any decision node to be
behaviour in the subtree starting at that particular node.
That brings me to the third and most important postulate, which I now call
consequentialism. This is like von Neumann and Morgensterns concept of
normal form invariance. It requires that there should be a consequence choice
function C dened on the domain of non-empty (and nite) subsets of Y with
the property that, whenever a decision tree T confronts the agent with the
feasible set of consequences F , behaviour in the tree T should limit the range
of resultant consequences to the choice set C(F ) F . In particular, whenever
two trees T and T
have the same feasible set F , behaviour in those trees should
give rise to the same consequence choice set C(F ). It is this condition that is
violated by both naive and sophisticated choice in the potential addict example.
To see this, note that in a dierent decision tree allowing the potential addict to
choose b in one stage perhaps by means of an infallible commitment device
rational behaviour would lead to the ex ante best consequence b instead of to
a or c. The same consequentialist condition is violated in the second example of
Section 5. This illustrates the diculty with the Pareto criterion or any other
incomplete preference relation.

7 Consequentialism Implies Ordinality

Recall that the three consequentialist conditions of Section 6 are unrestricted
domain, dynamic consistency, and consequentialism itself. Together, these three
conditions imply that behaviour in dierent feasible sets must generate conse-
quences that reveal a (complete and transitive) preference ordering over the
consequence domain Y . In other words, consequentialist rationality implies
ordinal rationality. In Hammond (1977), there is a proof based on Arrows
(1959) characterization of ordinal choice functions. But in more recent work I
have preferred to use simple direct proofs of this fact. Here is an outline of the
argument.
Consider any non-empty nite feasible set F Y . Now, there certainly
exists a one-stage decision tree T (F ) in which the only decision node is the
initial node, and the agents behaviour gives rise immediately to just one of the
consequences in F . It follows that for every non-empty nite F Y , the choice
set C(F ) is non-empty. Next, suppose that a belongs to the choice set C(F ).
Then a must be a possible consequence of behaviour in T (F ). But also, given
any other consequence b F , there is a dierent decision tree T with two stages,
as illustrated in Figure 3. The rst stage of the tree T occurs at initial node
n0 , where the agent chooses either a consequence y F \ { a, b }, or else goes on
to a second stage. This occurs at decision node n1 , where the choice is between
the consequences a and b.
In tree T , consequentialism requires option a to be one possible result of
the agents behaviour. Therefore, that behaviour must include the possibility
of moving from node n0 to n1 . Also, a must be a possible consequence of

9
 s a

n1



n0



P s
PP b
PP
PP
PP
PP s 
PP
PP y (y F \ { a, b })

PP
PPs

Figure 3: A two-stage decision tree T to demonstrate ordinality

behaviour in the subtree T (n1 ) beginning at node n1 , in which the feasible set
of consequences is { a, b }. Therefore a C({ a, b }).
On the other hand, if a C(F ) and b C({ a, b }), then b is also a possible
consequence of behaviour in the subtree T (n1 ). It follows that the move from
n1 to b is possible behaviour in T (n1 ) and so, by dynamic consistency, in T as
well. But it has been seen already that the agent may move from node n0 to n1
in tree T . Hence, b is a possible consequence of behaviour in the whole tree T ,
with feasible set F . That is, b C(F ).
Let us now dene the revealed preference relation R on Y by x R y i
x C({ x, y }). Because C({ x, y }) is non-empty for every pair set { x, y }, it
follows that the relation R is complete. Now, in the previous two paragraphs
it was proved that a R b whenever both a C(F ) and b F , and also that
b C(F ) whenever a C(F ), b F , and b R a. It is then very easy to show
that, for any nite set F Y , one has

C(F ) = { a F | b F = a R b }

So behaviour must reveal a choice function C that maximizes R over each nite
feasible set F Y .
Finally, to show that R is transitive, consider what happens when F =
{ a, b, c } and a R b, b R c. The three possibilities are:

1. a C(F ), in which case a R c because c F ;

2. b C(F ), in which case a C(F ) because a R b, so case 1 applies;
3. c C(F ), in which case b C(F ) because b R c, so case 2 applies.

Thus, a R c in every case, so R is indeed transitive.

Equally important is the fact that the ordinality property gives a complete
characterization of consequentialist behaviour. That is, given any preference
ordering on Y , there exists a behavioural rule dened for all nite decision
trees which satises dynamic consistency and consequentialism, while revealing
that particular preference ordering. This can be proved by a suitable backward

10
induction argument, like that involved in applying the principle of optimality in
dynamic programming.1 So consequentialism, just like ordinality, fails to limit
behaviour beyond the existence of some preference ordering. Behaviour which
maximizes a very unreasonable preference ordering will still be consequentialist,
even though it seems clearly irrational.

8 Consequentialism and Objectively Expected

Utility
As remarked already, the three consequentialist axioms of Section 7 imply noth-
ing more than the existence of a preference ordering in riskless decision trees.
But suppose that risk is introduced through chance nodes at which a move oc-
curs according to a specied or objective probability law. The consequences
of possible behaviour become simple probability distributions in the space (Y )
of lotteries with the property that (y) > 0 only for y in a nite support.
That is, consequences become risky.
The classical model of decision-making under risk requires behaviour to
result inrisky consequences (Y ) which maximize the expected value
IE v := yY (y) v(y) w.r.t. of what has come to be called a von Neumann
Morgenstern utility function (or NMUF) v : Y . Or, to be more exact,
a unique cardinal equivalence class of such functions, with all possible ratios
v(a) v(c)
(a, b, c Y ) of non-zero utility dierences uniquely dened and
v(b) v(c)
equal to constant marginal rates of substitution between shifts in probability
from c to a and shifts in probability from c to b. Nevertheless, in practice there
appear to be many systematic deviations from expected utility maximizing be-
haviour, such as those which Allais (1953, 1979a, b, 1987) and many successors
have noticed in various experiments. In response to these widely observed empir-
ical regularities, much attention has recently been paid to non-expected utility
models of decision-making under risk.
As I have pointed out elsewhere (Hammond, 1988b), non-expected utility
maximizers are liable to be essentially inconsistent in the same way as the po-
tential addict of Section 5. Indeed, there is even a risk of abandoning a project
that was initially judged to be benecial. To see this, consider the decision tree
illustrated in Figure 4, with four consequence lotteries , , , (Y ). At
the initial decision node n0 , the agent is confronted with the choice between
obtaining consequence for sure, or of going on to the chance node n1 . There
nature selects consequence with probability 1 , but with probability the
agent is taken to decision node n2 where the choice is between and .
Let v be any NMUF dened on Y . Then for any 0 < 1, the respective
expected utilities of lotteries , , must satisfy

IE v + (1 ) IE v IE v + (1 ) IE v IE v IE v
1 For details, see my forthcoming chapter Objective Expected Utility: A Consequentialist

Perspective to appear in the Handbook of Utility Theory (Kluwer Academic Publishers).

11
 s

n2



n1
s

n0



1

s



s

Figure 4: A decision tree to demonstrate essential inconsistency

But the expressions in the left hand inequality are the expected utilities of the
compound lotteries + (1 ) and + (1 ) respectively. So the
expected utility model implies the independence axiom (I) requiring that

+ (1 ) R + (1 ) R

As is well known, any continuous preference ordering on (Y ) satisfying (I)

can be represented by the expected value of each NMUF in a unique cardinal
equivalence class. So any non-expected utility maximizing agent with continuous
preferences must violate axiom (I).
Now, one way of violating the axiom (I) strictly would be to have

+ (1 ) P + (1 ) but P

Then, given continuous preferences on (Y ), there must exist (Y ) for

which
+ (1 ) P P + (1 ) and P
Yet these preferences almost reproduce those of the potential addict example.
At node n0 , the best option is +(1) . So a naive agent is likely to choose
node n1 , intending to continue on to node if given the opportunity. But if
node n2 is reached, the agent prefers to and so chooses . At node n0 , the
anticipated outcome of this choice is the lottery + (1 ) , which is worse
than either of the two alternative lotteries + (1 ) and faced by the
agent when at n0 . Like the sophisticated agent in the potential addict example,
the agent would do better to understand that is going to be made unavailable
by the agents own future behaviour. For this reason, should be chosen instead
of facing the risk of choosing over and so achieving + (1 ) . The
agent would do better still, of course, by nding some precommitment device
that collapses the decision tree and allows + (1 ) to be chosen without
any chance of replacing by .
Any such essential inconsistency will be avoided if and only if prescribed
behaviour at n0 in the subtree beginning at n1 is the same as in the subtree
beginning at n2 . But this requires precisely that axiom (I) be satisfed.
In this extended framework, what are the implications of the three con-
sequentialist axioms introduced in Section 7 namely, unrestricted domain,

12
dynamic consistency, and consequentialism itself? Somewhat surprisingly, they
lead to the absurd conclusion that there should be universal indierence, with
behaviour at each decision node left entirely arbitrary. Obviously, as before
the axioms imply the existence of a preference ordering. The trouble is that
axiom (I) has to be satised by this ordering even when = 0. So, for all
, , (Y ), one has R R . Since the left hand side of this
logical equivalence is a tautology, the right hand side must be true for all pairs
, (Y ), so there is indeed universal indierence.
There are three possible escapes from this ridiculous impasse. The rst
is to disregard those parts of a decision tree that can be reached only with
probability zero by pruning them o. The unrestricted domain of nite decision
trees is replaced by an almost unrestricted domain, excluding trees having
any zero probability move at some chance node. This restriction implies that
axiom (I) is only required to hold when 0 < < 1. Pruning o zero probability
branches in this way is eminently sensible in single-person decision theory. But
in multi-person games it leads to problems because it excludes consideration of
what would happen in the zero probability event that one player were to stray o
a presumed equilibrium path. In other words, it excludes all of the important
considerations that were captured so elegantly in Seltens (1965) concept of
subgame perfection.
A second escape is to weaken dynamic consistency and require it to hold
only in subtrees that are reached with positive probability. This is almost
dynamic consistency and allows + (1 ) I + (1 ) to be true when
= 0 even though P or P .
Elsewhere (Hammond 1994, 1997) I have begun to explore a third escape.
This involves arbitrarily small or innitesimal probabilities which are positive
yet less than any positive real number see also McLennan (1989) and Blume,
Brandenburger and Dekel (1991a, b). Any such probability can be regarded as
a suitable vanishing sequence of positive probabilities, in eect.
Given any of these three escapes, the consequentialist axioms imply that be-
haviour should have consequences which maximize a preference ordering R on
the space of consequence lotteries. Moreover, the ordering R should satisfy the
independence axiom (I). This is a complete characterization of consequentialist
behaviour because, given any such ordering on (Y ) that satises indepen-
dence, there exists behaviour satisfying the three consequentialist axioms which
generates R as the revealed preference ordering.
In particular, consequentialism does not imply expected utility maximization
because the preference ordering R could still embody discontinuities. To rule
these out, it is enough to invoke a requirement that, as probabilities at chance
nodes vary but the rest of the decision tree remains xed, including the conse-
quences at its terminal nodes, so the correspondence (or multi-valued mapping)
from probabilities at chance nodes to behaviour at decision nodes must have
a relatively closed graph. (Note that the graph cannot be closed because zero
probabilities are excluded.) With this fourth axiom of continuous behaviour,
one can prove that behaviour must have consequences in the form of lotteries

13
that maximize expected utility. Moreover, there is a unique cardinal equivalence
class of von NeumannMorgenstern utility functions.

9 Consequentialism and Unordered Events

Not all lack of certainty can be described by means of objective probabilities.
There can be uncertainty about unknown states of nature, instead of about risky
probabilistic prospects or lotteries. The deservedly famous decision theory of
Savage (1954) in particular sought to deal with this issue. He laid out seven
important postulates that are sucient to determine uniquely both an agents
subjective probabilities and also a cardinal equivalence class of von Neumann
Morgenstern utility functions.
Some of Savages most important postulates are logical implications of the
consequentialist axioms, when they are applied to nite decision trees with,
instead of chance nodes, natural nodes at which natures move helps determine
the state of the world or state of nature. Specically, suppose there is not
only a consequence domain Y but also a (nite) domain S of possible states of
the world aecting what consequence results from a given act. Then an event
E takes the form of any non-empty subset E S. For each such event E, there
must exist a contingent
 preference ordering RE dened on the Cartesian product
space Y := sE Ys of state contingent consequence functions y E = ys sE
E

mapping E into Y . Furthermore, the dierent conditional orderings must satisfy

Savages sure-thing principle requiring that, whenever E = E1 E2 where E1
and E2 are disjoint, and whenever aE1 , bE1 Y E1 , cE2 Y E2 , then

(aE1 , cE2 ) RE (bE1 , cE2 ) aE1 RE1 bE1

This is very like the independence axiom for lotteries with objective proba-
bilities, and it can be proved in essentially the same way that it is a logical
implication of the three consequentialist axioms.
Another of Savages postulates is that preferences are state independent.
This means that there must exist an ordering R on Y with the property that,
whenever s S and a, b Y , then a R{s} b a R b. In other words,
R{s} = R for all s S. This is not an implication of consequentialism, strictly
speaking. However, it makes sense to regard a decision tree in which only one
state of the world is possible as eectively identical to a corresponding decision
tree without any uncertainty. After all, the state of the world is known, as
also is the consequence of every strategy in the decision tree. If this accepted,
it makes sense to postulate that the consequences of behaviour in any such
decision tree should depend only on the feasible set of sure and certain conse-
quences in Y , independent of the state of the world. By repeated application of
the sure thing principle enunciated above, it is then easy to show that, when-
ever E S is an event and a1E , b1E denote the two constant state contingent
consequence functions satisfying as = a Y and bs = b Y for all s E, then


a1E RE b1E a R{s } b for all s
S. This fact is useful in the argument
of the next paragraph.

14
 Now, one of Savages most important postulates is that there be an ordering
of events from most to least likely. In particular, whenever E1 , E2 S are
disjoint events, it is natural to say that E1 is at least as likely as E2 provided
that, whenever a P b, then (a1E1 , b1E2 ) RE1 E2 (b1E1 , a1E2 ). In other words,
the agent weakly prefers the relatively favourable outcome a to occur in E1 and
the relatively unfavourable outcome b to occur in E2 , rather than the other way
around. What Savage assumed, following earlier work on qualitative probabil-
ity by Keynes (1921), Ramsey (1926), de Finetti (1937) and others, was that
this at least as likely relation is well dened. This is clearly necessary for the
existence of subjective probabilities, which must attach higher probability to a
more likely event. Yet the ordering of events condition is not an implication
of consequentialism, even when state independence is imposed and there is an
abundance of possible states. In fact, the consequentialist axioms imply no more
than the existence of dierent contingent preference orderings RE satisfying the
sure thing principle.
To conrm this, it is enough to exhibit a family of contingent preference
orderings that fails to induce an ordering of events despite satisfying the sure
thing principle. To do so, let S = { s1 , s2 } and Y = { a, b, c }. Then dene the
state independent utility function v : Y  so that:

v(a) = 1; v(b) = 0; v(c) = 1. (1)

Now consider the preference ordering on Y S induced by the specic additive

utility function
U S (y S ) = 1 (v(ys1 )) + 2 (v(ys2 )) (2)
where 1 and 2 are increasing functions satisfying

1 (1) = 2, 1 (0) = 0, 1 (1) = 1

(3)
2 (1) = 1, 2 (0) = 0, 2 (1) = 2

Suppose now that the two contingent orderings on Ys1 and Ys2 are represented
by the utility functions 1 (v(ys1 )) and 2 (v(ys2 )) respectively. Because (2) has
an additive form, the sure thing principle is evidently satised. Moreover, the
preferences on Ys1 , Ys2 are even state independent, as are those on the set
Y 1S := { (ys1 , ys2 ) Ys1 Ys2 | ys1 = ys2 }, since all are respresented by the
same utility function v(y). Nevertheless

U S (a, b) = 2, U S (b, a) = 1
U S (b, c) = 2, U S (c, b) = 1

So the agents behaviour reveals a preference for winning a in state s1 to winning

it in state s2 , when the alternative losing outcome is b. On the other hand, it
also reveals a preference for winning b in state s2 to winning it in state s1 , when
the alternative losing outcome is c. Hence, there is no induced ordering of the
events {s1 } and {s2 }.
Savage, of course, introduced other postulates whose eect is to ensure a
rather rich set of states. Adding such postulates, however, in general will not

15
induce an ordering of events. To see this, suppose that S is the entire interval
[0, 1] of the real line instead of just the doubleton { s1 , s2 }. Instead of the
additive utility function (2), consider the integral
 1/2  1
S (y S ) =
U 1 (v(y(s)))ds + 2 (v(y(s)))ds
0 1/2

with v given by (1) and 1 , 2 by (3). Also, so that the integrals are well
dened, y S should be a measurable function from S to Y , in the sense that the
set { s S | y(s) = y} is measurable for all y Y . Then the particular CCF
y S = a 1[0, 12 ] , b 1( 12 ,1] with

a if s [0, 12 ]
y(s) =
b if s ( 12 , 1]
 
is preferred to the lottery represented by b 1[0, 12 ] , a 1( 12 ,1] in the same notation.
   
But c 1[0, 12 ] , b 1( 12 ,1] is preferred to b 1[0, 12 ] , c 1( 12 ,1] . So there is no induced
likelihood ordering of the two events [0, 12 ] and ( 12 , 1]. In fact, it is easy to
conrm that this example satises Savages postulates P1P3 and P5P7; only
the ordering of events postulate P4 is violated.

10 Subjective Expected Utility

The subjective expected utility model can be given a consequentialist justica-
tion, however. Doing so requires extending the domain of nite decision trees
so that, as well as decision nodes and terminal nodes, one can have both the
chance nodes of Section 8, and also the natural nodes of Section 9. Then the
consequences of behaviour become lotteries over state contingent consequence
functions. On the space of such lotteries the three consequentialist axioms im-
ply the existence of contingent preference orderings satisfying both the indepen-
dence axiom and the sure thing principle. Even so, and even if one also imposes
the continuous behaviour axiom of Section 8 and the state independence con-
dition discussed in Section 9, there still remain several possible deviations from
subjective expected utility. For details, see Hammond (1988a).
However, a decision tree which has both chance and natural nodes is al-
most like a three-person game in extensive form. The players are the agent, of
course, but also chance, who moves according to objectively specied prob-
ability laws, and nature. Applying the idea of consequentialist normal form
invariance to such a tree helps justify a key axiom of an alternative approach to
subjective probability due to Anscombe and Aumann (1963). This key axiom
is called Reversal of Order, or (RO). To state this condition, let = i ki=1
be any probability distribution or roulette lottery over the set of integers
i { 1, 2, . . . , k }, and let Si (Y S ) (i = 1, 2, . . . , k) be any collection of k
lotteries, with is (y) as the marginal probability of obtaining consequence y in

16
 k
each state s E. Let i=1 i Si (y S ) be the compound lottery in which the
roulette lottery is resolved before the horse
k lottery that determines which
s E occurs. Then condition (RO) requires i=1 i Si (y S ) to be indierent to
the alternative compound lottery in which the horse
k lottery is resolved rst, and
its outcome s S determines which of the list  i=1 i is (ys ) sS of marginal
roulette lotteries occurs next.

T T

n1 s x1s 1 s x

n
1
s 1s
s
1 s

s
n0
x1s n
0
s 2 x2s


s x2s 1 s x


2
s s
1s
s

n2 s x2s

n2 2 x2s
Figure 5: Decision trees T and T
(when k = 2 and E = { s, s })
This condition can also be given a consequentialist justication by showing
how it is implied by a hypothesis I call consequentialist normal formal invari-
ance. To illustrate this, consider two decision trees T and T
described as
follows. Tree T begins with the chance node n0 , which is succeeded by the set
of natural nodes N+1 (n0 ) = { ni | i = 1, 2, . . . , k }. The transition probabilities
are (ni |n0 ) = i (i = 1, 2, . . . , k). Each ni is succeeded by the set of terminal
nodes N+1 (ni ) = { xis | s E }. The consequences are assumed to be given by
(xis ) = is (Y ) (s E). Tree T is illustrated in the left half of Figure 5
for the case when k = 2 and E = { s, s }.
On the other hand, tree T
begins with the natural node n
0 , whose successors


form the set N+1 (n
0 ) = { n
s | s E }. Then each n
s is a chance node whose


successors form the set N+1 (n
s ) = { x
is | i = 1, 2, . . . , k } of terminal nodes. The
transition probabilities are
(x
is |n
s ) = i (i = 1, 2, . . . , k). The consequences
are assumed to be given by
(x
is ) = is (Y ) (s E). Tree T
is illustrated
in the right half of Figure 5, again for the case when k = 2 and E = { s, s }.
Both trees represent a three-person extensive game between chance, nature,
and the agent, who actually has no decision to make. In tree T it is natural to
assume that N+1 (n0 ) is a single information set for nature. Similarly, in tree T



it is natural to assume that N+1 (n
0 ) is a single information set for chance. Then
the extensive form games represented by the two trees will have identical normal
forms, in which the agent has only one strategy, whereas chances strategies
are indexed by i { 1, 2, . . . , k } and natures strategies are indexed by s
E. Furthermore, in either extensive form game, when chance chooses i and
nature chooses s, the consequence is the lottery is (Y ). In fact, the
consequentialist normal form invariance condition is that trees like T and T

with identical three-person normal forms should be regarded as giving rise to
equivalent feasible sets, and that behaviour should generate equivalent choice
sets of consequences in each case. This is a natural extension of the normal
form invariance hypothesis mentioned previously.

17
 Evidently,
k in tree TE the sole feasible consequence k available to the decision


maker is i=1 i E i (y ), whereas in tree T it is  i=1 i is (ys ) sE . Conse-
quentialist normal form invariance requires these to be regarded as equivalent.
But this is precisely condition (RO).
With this reversal of order condition added to the three consequentialist
axioms, plus state independence and continuity, it follows that behaviour must
maximize the expected value of each NMUF in a unique cardinal equivalence
class, with subjective probabilities applied to the dierent possible moves by
nature. Moreover, all these subjective probabilities must be positive. In par-
ticular, no event can be null according to the criterion specied by Savage.
Furthermore, as the agent moves through the decision tree, these subjective
probabilities must be revised according to Bayes Rule. This is a strengthening
of the standard subective expected utility (or SEU) model; accordingly, it will
be called the SEU* model.
As in previous sections, it can also be shown that the SEU* model is the
only implication of these axioms. Nothing yet rules out absurd utility func-
tions. Or, in the case of subjective probabilities, absurd beliefs reected in
these probabilities. We are still not past rationality as structural consistency.
But at least consequentialism, with its focus on what is reasonable behaviour
in non-trivial decision trees, seems to oer a stronger defence of the standard
structural consistency criterion and the associated SEU model.
That concludes the discussion of normative single-person decision theory in
this paper. It is time to move on to two branches of multi-person decision
theory. The rst is social choice theory, which is concerned with nding a
common objective for a group of heterogeneous individuals. And the second is
game theory, which studies how dierent individuals interact in their behaviour.

11 Social Choice Theory

Most irrational behaviour by ordinary people has little eect outside a small cir-
cle of family, friends, workmates, etc. Mistakes by professionals such as doctors
can be somewhat more devastating, and lead to unnecessary loss of property,
life or limb. So can reckless driving by any of us. Of course, Akerlof and Yellen
(1985a, b) argued that even small mistakes, when made by many people simul-
taneously, can have surprisingly large adverse consequences. Nevertheless, the
irrational behaviour that has the largest eects on the most people comes from
political leaders, public ocials, etc. Even in contemporary Europe, their mis-
takes and/or criminal behaviour can still destroy nation states. It seems obvious
that we should expect those granted substantial power over their colleagues or
fellow citizens to meet higher standards of rationality.
Now, in my view social choice theory should be about specifying suitable
objectives for public ocials and others responsible for major decisions aecting
large numbers of individuals. In particular, it should specify objectives for
economic policy. Accordingly, the structural consistency conditions discussed
in previous sections seem even more appropriate for social choice than they are

18
for individual behaviour. This suggests that the structure of the social decision
tree should not aect how the consequences of public decisions emerge from
the set of feasible consequences. So the consequentialist axioms apply, and
imply the existence of a social ordering. When there is risk, it is appropriate
to maximize the expected value of a cardinal von NeumannMorgenstern social
welfare function (or NMSWF). And to attach positive subjective probabilities to
states of nature which do not have specied objective probabilities, so reaching
the SEU* model.
However, the NMSWF W and its expected value IEW should reect the
interests of the individuals who are aected by the decisions under consideration.
In principle, this could even be the entire world population, together with future
generations yet to be born. In order to model these interests, it helps to adapt
a device used by Foley (1970) and Milleron (1972) to treat public goods, and
imagine for a moment that each individual could have their own separate world
to live in.
Let I denote the set of all individuals. For each i I, the choice of world for
individual i, taking account of only is personal interests, is a degenerate social
choice problem. Applying the consequentialist axioms and their relatives to the
class of all degenerate decision trees involved in choosing a world for i, we are
able to infer that the interests of each individual i should be represented by the
expected value of their own von NeumannMorgenstern individual welfare or
utility function wi . That is, if we were choosing a world for i alone, it would be
appropriate to maximize IEwi .
Now consider any social decision tree which happens to be one individual
is personal decision tree in the sense that, although one is choosing a separate
world for each individual, in fact any decision that is made aects only is world,
leaving all other individuals worlds completely unaected. For such decisions,
it seems reasonable to insist that the social objective IEW should collapse to the
individual objective IEwi , in the sense of representing the same social preference
ordering. But then it is easy to prove that, whenever two lotteries over possible
worlds for all individuals are equally attractive for all those individuals, society
must be indierent between them. From these arguments, the strict Pareto
condition follows. That is, if all individuals are no worse o, then so is society;
also, if any individual is better o, then so is society unless some other individual
is worse o. In particular, if all individuals are equally well o, then so is society;
this is the Pareto indierence condition.
Given these Pareto properties, and the fact that each individuals risky world
can be chosen separately and independently, it follows that the social welfare
function is some positively weighted sum of all individuals welfare functions. In
fact, this was rst proved by Harsanyi (1955). By suitable normalization of the
cardinal individual welfare functions, these weights can all be made equal to 1.
Also, where decisions aect the set of individuals who come into existence, one
can presume that individuals who will never come into existence have no interest
in what world is created for them, and so have constant utility in all such worlds
of permanent non-existence. Then we can normalize individual utilities so that

19
this constant utility is zero. Formally, then, the social welfare function takes
the classical utilitarian form, being equal to the sum of all individual utilities.
This objective is for an articial problem where all individuals can have
their own separate and independent risky worlds. But the natural constraint
that we must all share the same world, and so that all individuals risks must be
perfectly correlated, is merely a restriction on what decision trees are relevant.
The form of the objective is not aected.
The social welfare objective specied above is formally the same as for clas-
sical utilitarianism. But the utilities being added have quite a dierent in-
terpretation. Expected individual utility is dened in a way that makes it an
appropriate objective in decision trees aecting only one individual. And it can
be shown rather easily that each utility ratio amounts to a constant marginal
rate of substitution between the numbers of individuals in two particular sit-
uations. In particular, it is better to change society in order to have more
individuals with higher utility, and fewer with lower utility. It is also better to
have more individuals with positive utility, but this should be interpreted care-
fully because the zero level of utility is dened so that society is better o with
more individuals who have positive utility, but worse o with more individuals
who have negative utility. For more details, see Hammond (1991, 1996).
Once again, by thinking carefully about what is appropriate behaviour in
decision trees, some surprisingly strong conclusions have emerged.

12 Some Special Two-Person Games

Even two-person games are much more complicated than single-person decision
trees. Rational behaviour for each player usually depends on what behaviour is
rational for the other player. The only exception appears to be in games like
Prisoners Dilemma, or nite repetitions of it, in which each player has a unique
strategy that strictly dominates all others i.e., is uniquely best regardless
of what strategy the other player chooses. Yet even for Prisoners Dilemma,
some people have argued that, at least in special circumstances, it is rational to
play the dominated strategy of cooperating instead of the dominant strategy of
defecting. For example, suppose that Prisoners Dilemma is going to be played
between two computer programs. Then Howard (1988) showed how to write a
program which checks to see if its opponent is an exact copy of itself. Such a
program could then go on to cooperate if and only if it is playing against such
a copy.
Another class of relatively simple two-person games are those in which there
is perfect and complete information, and the game is bounded in length. An
example of such a game is chess, and indeed the rst modern mathematical work
on game theory was Zermelos (1912) pioneering analysis of chess. He proved
that, with best play by both White and Black, chess has a denite result a
win for White, a draw, or a win for Black. Which of these is right depends on
a nite calculation, but one which is so complicated that it may not be possible
ever to determine what is the result of a perfectly played game of chess. Clearly,

20
normative models of behaviour lose their practical signicance when the decision
problem is too complicated ever to be analysed fully.
The procedure for resolving chess, in theory but not in practice, is backward
induction. This works backwards from terminal positions in which the game
is already over because there is checkmate, stalemate, a draw by repetition or
perpetual check, or a draw because each player has made 50 moves without
capturing a piece or moving a pawn. Then the backward induction procedure
evaluates all positions where one player can force checkmate in one move, or else
can only avoid being checkmated by moving immediately to a terminal drawn
position. The procedure goes on to evaluate positions which, given best play
by both sides, require at most two moves to reach a terminal position, then at
most three, etc. In principle, all possible positions get evaluated in this way,
including the starting position.
The rules of chess give the two players strictly opposing objectives, so the
normative relevance of backward induction is uncontroversial. Indeed, it is ap-
plied in practice in order to evaluate simple chess endgames. But backward
induction is controversial in other important games like the Centipede con-
sidered by Rosenthal (1981) and many later writers. This game is illustrated in
Figure 6 (see also Osborne and Rubinstein, 1994).

P1 P2 P1 P2 P1 P2 P1 P2
s s s s s s s an
s bn s
n+1,n
a0 b0 a1 b1 ak bk an1 bn1

s s s s s s s s
1,0 0,2 2,1 1,3 k+1,k k,k+2 n,n1 n1,n+1

Figure 6: Rosenthals Centipede Game, Modied

There are two players labelled P1 and P2 with respective strategy sets

S1 = { ai | i = 0, 1, . . . , n }, S2 = { bj | j = 0, 1, . . . , n }

If P1 chooses ai S1 , then unless P2 ends the game beforehand, P1 moves

across in the tree exactly i successive times before moving down. Similarly, if
P2 chooses bj S2 , then unless P1 ends the game beforehand, P2 moves across
exactly j successive times before moving down. The game ends immediately if
either player moves down. If i = j = n, it ends anyway after P2 has chosen bn .
Outside this case, if the two players choose (ai , bj ) where i = k j, then the
game ends after they have both moved across k times before P1 moves down.
But if i > k = j, then P1 moves across k + 1 times but P2 only k times before
moving down to end the game. Hence, the two players respective payos are
 
i + 1 if i j i if i j
v1 (ai , bj ) = and v2 (ai , bj ) =
j if i > j j + 2 if i > j
It is now easy to see that whenever P2 is given the opportunity to choose bn ,
it would be unreasonable to do so because bn1 is available and yields a higher
payo. But once bn has been eliminated, whenever P1 is given the opportunity

21
to choose an , it would be unreasonable to do so because an1 is available and
yields a higher payo. Then, for k = n 1, . . . , 1, once an , . . . , ak+1 have all
been eliminated, each successive bk is worse than the alternative bk1 . Similarly,
for k = n 1, . . . , 1, once bn , . . . , bk have all been eliminated, each successive
ak is worse than the alternative ak1 . Thus, the backward induction procedure
successively deletes all strategies until the only ones remaining are b0 for P2
and a0 for P1 . The strategy prole (a0 , b0 ) is one Nash equilibrium, of course,
though there are others as well. The tree collapses to the single branch a0 .
Backward induction suggests that this is the only part of the tree which needs
to be analysed.
Starting with Rosenthal (1981) and Binmore (1987), several game theorists
have found this backward induction argument to be unconvincing, for the fol-
lowing reason. Suppose P1 were unexpectedly faced with the opportunity to
play ak after all, because neither player has yet played down, and in fact each
player has played across k times already. Backward induction applied to the
remaining subtree leads to the conclusion that P2 , if given the move, will play
bk next time, so P1 should play ak . Yet P2 has already played across k times,
whereas backward induction implies that P2 should move across whenever there
is a move to make. So, as Binmore in particular argues most persuasively, if k
is large enough, P1 has every reason to doubt whether the backward induction
argument applies to P2 s behaviour after all. Furthermore, if n k is also large,
there may be much to gain, and at most 1 unit of payo to lose, from allowing
the game to continue by moving across instead of playing ak .
Of course, P2 can then apply a similar argument when faced with the choice
between bk and continuing the game. Also, P1 should understand how moving
across once more instead of playing ak will reinforce P2 s doubt about whether
the backward induction argument applies to P1 , and so make it more likely
that P2 will decline to play bk . This strengthens P1 s reasons for not playing
ak . Similar reasoning then suggests that P2 should not play bk1 , that P1
should not play ak1 , etc. In the end, it may be sensible for P1 not to play
a0 , for P2 not to play b0 , etc. Indeed, there are some obvious close parallels
between this argument and that of Kreps, Milgrom, Roberts, and Wilson (1982)
for players rst to cooperate and then play tit-for-tat in the early stages of a
nitely repeated Prisoners Dilemma. Or for a chain-store to play tough in
the paradox due to Selten (1978).

13 Perfect Recall
One major rationality assumption in orthodox game theory is perfect recall.
This requires that, in an extensive form game tree, all the branches passing
through a particular information set of any player must pass through an iden-
tical sequence of that players previous information sets; in addition, at each
preceding information set of that player, there must be exactly one strategy al-
lowing the later information set to be reached. The interpretation is that players
remember all the information that was available previously, and also whatever

22
moves they chose at their own previous information sets. Because agents with
ner information partitions are able to make better decisions, it seems obvi-
ous that players in a game should strive for perfect recall as a rather obvious
normative standard.
Actually, for some years, I thought that the assumption of perfect recall could
be made without loss of generality. After all, a player who forgets something
can be treated as a dierent player. So there seems no harm in treating each
information set as belonging to a dierent local player, and replacing each
original player by a team of local players who share a common objective but
may not be able to share their information. For example, it seems not to
matter much whether one regards the game of bridge like von Neumann and
Morgenstern did, as a game between two players, NorthSouth and EastWest,
each with imperfect recall. Or instead as a four person game, with North, South,
East and West treated as separate players who form two opposing teams.
Piccione and Rubinstein (1994) point to an interesting counter example,
however. When a player has imperfect recall, the same information set can be
encountered more than once along some branches of the game tree. In this case,
imperfect recall plays an essential role, and even in a single person game it is
possible for a mixed strategy to be better than any pure strategy. However, this
is a rather exceptional situation in which the player loses count of how many
times a particular information set has been visited. Now, players who can keep
count of how often they visit each of their information sets must actually have
information sets which can only be visited at most once. For such players, it is
possible to treat imperfect recall by introducing enough local players.

14 Nash Equilibrium
Two-person games of perfect and complete information like chess or the Cen-
tipede can be solved by backward induction. They are not typical, however.
Generally, players must move simultaneously, or in ignorance of the others pre-
vious moves. They would like to know these moves so that they could react
appropriately, but they do not and cannot know.
An important class of games are those like chess with two players who have
strictly opposing objectives. What one wins the other loses. Hence the name
zero-sum, though this terminology does not t in well with modern ordinal
or even cardinal concepts of utility which cannot be added across people. For
such games, Borel (1924) suggested that mixed or randomized strategies could
be important, as when a poker player chooses at random whether to blu or
not. A little later, von Neumann (1928) proved that allowing two players in a
zero-sum game to choose mixed strategies would give each player a determinate
maximin expected payo. This can be found by maximizing with respect to
his own strategy the payo that is minimized with respect to the opponents
strategy.
Meanwhile, Morgenstern (1935) had realized that economic forecasting was
dicult or impossible unless one could somehow determine agents expectations

23
concerning the future, including their expectations of other agents behaviour,
which would depend on these other agents expectations, etc. The two-person
zero sum game theory which von Neumann had developed must have seemed at
the time a way of resolving the resulting innite regress. So they got together to
write von Neumann and Morgenstern (1944, 1953) one of the most important
and inuential scientic books of all time. However, only their analysis of two-
person zero-sum games is really satisfactory; indeed, the last two-thirds of their
book (apart from the general concept of a stable solution set, and the appendix
on expected utility) receives little attention these days, and is probably only of
historical interest.
The next major step in game theory is due to Nash (1950, 1951) who was able
to show that any n-person game would have an equilibrium in mixed strategies,
with each players mixed strategy maximizing their own expected utility given
all the other players mixed strategies. The concept of Nash equilibrium is one
of the most widely used in economics. Yet it is not without its conceptual
diculties, as Luce and Raia (1957) in particular pointed out rather early on.
If a game has a unique Nash equilibrium, both players might be expected to
nd and play it. If two players 1 and 2 have sets of equilibrium strategies E1
and E2 with the property that any pair (s1 , s2 ) E1 E2 is a Nash equilibrium,
then the equilibria are said to be exchangeable. In this case, any equilibrium
s1 E1 is a best response to any equilibrium s2 E2 , and vice versa. In
particular, whenever player 2 chooses any s2 E2 , all equilibrium strategies
s1 E1 give player 1 exactly the same expected payo, and vice versa. So
each player i should be content to play an arbitrary strategy si Ei , and
thereby reach one of the multiple equilibria. But there remain many games
like Battle of the Sexes with non-exchangeable multiple Nash equilibria, even
in pure strategies. That is, the equilibrium set might well contain at least two
strategy pairs (s1 , s2 ), (s
1 , s
2 ) for which s1 must be a best response to s2 , by
denition, but may not be a best response to s
2 . Then, to be sure of reaching
a Nash equilibrium, player 1 needs to know which equilibrium strategy player 2
is likely to choose, but cannot. Similarly for player 2.

15 Rationalizability
This problem of non-exchangeable strategies went unresolved for about 25 years,
until Bernheim (1984) and Pearce (1984) included independent work on ratio-
nalizable strategies in their respective Ph.D. theses, and had papers published in
Econometrica. One of the key steps in the Bernheim/Pearce revolution was their
replacement of the objective probabilities that are specied in mixed strategies
by subjective probabilities which each player attaches to the others strategies.
Previously, Harsanyi (19678) had discussed how players might attach subjec-
tive probabilities to each others types in a game of incomplete information.
Also, related ideas can be found in the somewhat obscure work of Armbruster
and B oge (1979) and of B oge and Eisele (1979). But it was Bernheim and

24
Pearce, followed shortly by Tan and Werlang (1988), who developed a useable
new game theory.
In fact, rationalizable strategies are those that are optimal given rationaliz-
able beliefs in the form of subjective probability distributions attached to other
players strategies. And rationalizable beliefs are those attaching probability
1 to the event that all other players will choose some rationalizable strategies.
Bernheim and Pearce showed how one could apply this obviously circular deni-
tion rather easily. Indeed, rationalizable strategies in two-person games could be
characterized as all those that survive iterative deletion of strictly dominated
strategies for both players. There is a similar characterization for n-person
games provided that rationalizable beliefs allow other players strategies to be
correlated whenever necessary.
Though any strategy appearing in Nash equilibrium is certainly rationaliz-
able, there can be many other rationalizable strategies besides. So, in an era
when many game theorists were applying ideas due to Selten (1965) or Kreps
and Wilson (1982) in order to rene the set of Nash equilibria, it is perhaps
surprising that the Bernheim/Pearce coarsening of the equilibrium set received
as much attention as it did. This may be because of the persuasiveness of their
approach, which simply and conclusively answers the question of what minimal
restrictions standard decision theory imposes on possible behaviour in a normal
form game.
Nevertheless, as Mariotti (1996) points out, there does remain a rather se-
rious aw in the arguments used by Bernheim and Pearce, as well as by Tan
and Werlang and others. After all, they assume that players have beliefs about
each others strategies that are represented by subjective probabilities, and that
players take decisions according to the SEU model described in Section 10.2 In
Sections 610, these assumptions were justied by axioms concerning behaviour
in an almost unrestricted domain of decision trees. In fact subjective probabili-
ties relate to the willingness of the agent to trade o good consequences against
bad in dierent states of the world in other words, to make particular bets.
Instead of variations in single-person decision trees, one might assume directly
that players have preferences over consequences. But then such preferences re-
ally have meaning only when players are faced with choices between pairs or
other subsets of consequences. In other words, in order to apply the axioms of
decision theory to games, and so give preferences decision-theoretic content, one
must allow variations in the set of strategies and their consequences which are
available to each player.
Now, in single-person decision theory, it makes sense to assume that nature
determines the uncertain state of the world unsystematically or haphazardly.
Then it also makes sense to assume that natures choice is unaected by varia-
tions in the agents decision tree or other representation of the decision problem.
2 In his work on correlated equilibrium, Aumann (1987) is more careful. He postulates a

space of extrinsic states, each of which is assumed to determine a unique strategy prole that
is suggested to all players before the game starts. Equilibrium occurs when individuals have
identical beliefs about the probabilities of all the dierent extrinsic states, and when for each
player it is optimal given these beliefs to follow the suggested strategy.

25
But in games, all players are assumed to choose strategies in their own best in-
terests. Then variations in one players strategy set aect the whole game in
ways which are likely to aect what strategies it is in these other players in-
terests to choose. For this reason, as Mariotti argues, it is far from evident
that one can legitimately treat other players purposefully chosen strategies in
the same way as moves determined haphazardly by nature. Above all, it is
not obvious that the SEU criterion can be applied to bets contingent on other
players strategies in the same way as to bets contingent on states of nature.
Nor is it obvious that, even if players are willing to accept bets in such a way,
they should treat the game itself as eectively a similar bet to which the SEU
criterion can be applied.
To overcome this serious diculty, Battigalli (1996) suggests that instead
of having the players themselves accept bets on strategy choices in the game,
one could consider a clone of each player whose behaviour in identical decision
problems will be exactly the same. This clone is regarded as entirely outside
the game, in the sense that the clones decisions have no eect whatsoever on
the consequences of the strategies chosen by any of the players participating
in the game. Then the clone is in a position to treat the game in the same
way as a small gambler who is not a jockey treats a horse race. That is, the
clone can regard the strategy choices of all the players in the game as states of
nature, with subjective probabilities attached. Also, the clone can be faced with
a decision problem that is eectively a copy of that faced in the game itself by
the corresponding original player. Because the clone will maximize subjective
expected utility in this copy of the game, the original player will do the same
in the original game.
In this way, by applying the axioms presented earlier in this essay in a
suitable context, one can justify the claim that each player should indeed use
the SEU model, just as most game theorists have always asserted. The same
axioms also justify the standard specication of a game in normal form, with
payo functions that are really NMUFs. There is the obvious qualication
that each players payo function or NMUF is only determined up to a unique
cardinal equivalence class. Clearly, this is unimportant. Note, however, that the
framework used here diers from that of B orgers (1993), whose assumptions do
not allow decision problems with objective probabilities to be considered, and
so yield only a unique ordinal equivalence class of utility functions.
Another requirement is that each player is SEU model include a unique
specication of appropriate subjective probabilities for other players strategy
proles in the game, as well as for any exogenous states of nature which may
be relevant. Failure to specify these probabilities leaves the description of the
players decision models fundamentally incomplete. Yet specifying them arbi-
trarily ignores the fact that, in the end, other players strategies are really not
like states of nature, because other players face their own individual decision
problems in the game, which they try to resolve rationally, at least to some ex-
tent. This tension is precisely the challenge that non-cooperative game theory
must meet.

26
 A more serious worry is that, as explained in Section 10, consequentialism
actually implies the SEU* model, in which each player is subjective probability
attached to any prole of other players strategies must be strictly positive.
This contradicts much of orthodox game theory, where each player is required
to attach zero probability to the event that one or more other players choose
strategies which are not their best responses. In particular, the probability of
any player choosing a strictly dominated strategy must be zero.
As mentioned in Section 8, one remedy might be to weaken dynamic con-
sistency to almost dynamic consistency. In extensive form games, however, this
approach forces separate consideration of every subgame of incomplete infor-
mation that a solution is presumed to reach only with zero probability. This
eectively violates the normal form invariance hypothesis, as enunciated by von
Neumann and Morgenstern (1944, 1953), which claims that it is enough to anal-
yse the normal or strategic form of any game. This invariance hypothesis, more-
over, especially applied in its consequentialist form, is really what justies the
subjective expected utility approach to decision making, which represents the
key assumption of modern game theory. The only satisfactory remedy appears
to be using innitesimal probabilities of the kind mentioned in Section 8.

16 Restrictions on Beliefs
The SEU* model, even within a general multi-person game, actually imposes
very few restrictions on agents beliefs. After all, it is a structural consistency
model of rational behaviour. About the only way in which an agent who uses
the SEU* model properly could ever be inconsistent is to reach some unforeseen
situation. So the only advice it seems to oer is to take every possibility into
account and accord everything a positive small probability. Be broad-minded,
in other words.
Excessive broad-mindedness seems clearly irrational, however. One should
not believe the Earth is at. One should believe the sun will rise tomorrow
(unless you are in either the Arctic or the Antarctic during winter). One should
expect most people who are young enough to learn from their obvious mistakes
most of the time. Earthquakes will occur in California, avalanches in the Alps,
accidents on the roads, and volcanic eruptions on Mount Vesuvius or Etna.
These are just instances of physical or psychological reality which we ignore at
our peril. The question of what it is reasonable to believe in such instances
represents the hard philosophical problem of induction. It has exercised many
great minds over several millenia without reaching any very good resolution that
I am aware of.
Nevertheless, several distinguished game theorists such as Harsanyi and Au-
mann have argued that players would have common probabilistic beliefs if they
pooled all their information. Equivalently, it as though all players started with
a common prior distribution which then became updated based on information
acquired by each player just before the game starts. In the case when this
common prior also concerns what strategies the dierent players in a game will

27
choose, they will then reach a Nash equilibrium or, if correlation is allowed, a
correlated equilibrium (see especially Bernheim, 1986). Similar ideas underline
the concept of rational expectations equilibrium that was made famous by Muth
and Lucas.
Where a game is played only once, it seems obvious that players will not
have enough time to learn what other players are likely to do. After all, in-
duction can only apply to situations that are repeated suciently often. Thus,
the hypothesis of rational expectations seems implausible except in games that
repeated many times, or where players in a one-shot game have reason to
believe that they all play strategies appropriate to a particular focal equi-
librium (Schelling, 1960). In fact, though I am certainly willing to admit that
players can and should learn from their past experiences if they are in a repeated
game, I have never been persuaded that most of the games economists should
be interested in really are repeated in the relevant technical sense, nor that one
should limit attention to rational expectations or common prior beliefs. So
the idea that agents should make use of all available information is a good one,
but too much is likely to remain hidden to allow equilibrium to be reached in
all but trivial games which, for instance, happen to have unique rationalizable
strategies.
Of course, abandoning equilibrium does much to weaken the predictive power
of game theory. After all, if a game has a unique equilibrium, that can be used
to determine each players strategy and also each players expectations of the
other players strategies. The shift from equilibrium to rationalizability typically
expands the set of strategies and of associated expectations which rationalize
those strategies. Then one may be forced to look for considerations beyond the
scope of traditional game theory in order to help predict players expectations
and their strategy choices.
Actually, even in traditional game theory, considerations that are excluded
from the traditional description of a game may be relevant to the choice of equi-
librium. An important example is forward induction, as discussed by Kohlberg
and Mertens (1986). This works most clearly when the rst player to move
has the chance to end the game immediately with a well dened consequence
or outside option, or alternatively of entering a subgame to be played with an-
other player. Typically, this subgame, like Battle of the Sexes, has multiple
non-exchangeable equilibria. Then the claim is that, if the rst player chooses
to enter this subgame, it can only be in the expectation of achieving an out-
come that is no worse than the outside option. In several interesting games,
this is enough to restrict the relevant equilibrium set in the subgame to a single
outcome. For further discussion, see van Damme (1989), Osborne (1990), Hillas
(1990), Hammond (1993), and especially Gul (1996).
Notice however that, if forward induction can be used to rene the equilib-
rium set within some subgame, perhaps it could also be used for an entire game.
After all, one or other players may have given up past outside options to enter
the game that is about to be played. So, in addition to the usual description of
a game in normal or extensive form, in some cases it is also important to know

28
what happened before the game started. Traditional game theory ignores all
such considerations.

17 Beyond Structural Consistency

The previous section was concerned with possible restrictions on beliefs that
transcend the usual SEU* model, which recommends only that everything which
might be relevant should be taken into account. It is also reasonable to impose
restrictions on preferences. Up to a point, more of the good things that life has
to oer should be preferred to less. Excessive risk of severely adverse outcomes
like death or mutilation should clearly be avoided. People who have any funds
in savings accounts should use them to pay o high interest credit card balances
as rapidly as possible. Many people should probably buy fewer lottery tickets
than they do, unless serious nancial distress is so imminent that the only hope
of escape is to win a large prize. Schoolchildren should be taught how to handle
nancial aairs, including how to manage pension accounts, mortgages, etc. Not
only schoolchildren, in fact. Those who are about to retire and go on a pension
often need advice from a source other than agents seeking a fat commission.
These rather vague thoughts illustrate the point that, even when choosing on
their own behalf, individuals may well benet from guidance in establishing what
their preferencess should be, as well as their beliefs about the likely consequences
of dierent possible actions. Full rationality is much more than mere structural
consistency. This is especially true if one believes that full rationality also
requires suitable attention to ethical values and respect for other people, as well
as animals, etc. Such considerations, and the judgement that some preferences
are rational while others are not, goes beyond traditional economics and into the
realm of philosophy. At this point, economists should recognize their limitations
and admit that they know little about many fundamentally important questions
concerning the rationality of dierent preferences. So, counting myself as an
economist instead of a philosopher, I shall pass on to the next topic.

18 Procedural Rationality versus Satiscing

The SEU* model recommends taking all possibilities into account, examining
every available decision for its likely consequences, and then maximizing. The
implicit methodology regards problems like playing chess perfectly as relatively
trivial. Such unbounded rationality is clearly far too exacting.
A little over 40 years ago, Simon (1955, 1957) realized this and went on to
introduce an alternative concept which he called procedural rationality. Trying
to maximize subjective expected utility, even when it is feasible, is often unnec-
essarily bundensome. Far better to satisce by nding an action that meets
an appropriate aspiration level.
Part of Simons theory, even its main message, makes eminent sense. The
unbounded rationality of the SEU model does need replacing with some less

29
onerous procedure for making boundedly rational decisions. However, the
satiscing approach does not seem satisfactory, especially as a normative stan-
dard. Satiscing might make sense if the only diculty were to nd the best
decision among a large number of possibilities being given due consideration.
Instead, it seems to me that a much more common problem arises because it is
very hard to analyse more than a very small number of options in any detail,
whereas it is rather easy in principle to rank those which have been analysed
thoroughly and so nd the best from this restricted set. In other words, agents
who face complicated decision or game trees are forced to include only very few
possibilities within any conceptual model they use to analyse the decision they
have to make. In other words, they bound their models.
Now, a bounded model is likely to be worth revising later on, when the agent
is closer to the time when an important decision has to be made, and history
has rendered irrelevant many of the contingencies which had been possible ex
ante. Of course changes in the agents model of the decision tree may aect the
consequences of behaviour. But if the agent never sees any need to revise his
bounded model, there is no reason why its structure should inuence behaviour
and its consequences. And even if the bounded model is likely to change, for
behaviour within whatever bounded model the agent chooses to construct, all
the standard axioms retain their normative appeal. In particular, they recom-
mend that, within the bounded model, the agent should choose the best possible
action. This seems quite dierent from Simons recommendation that the agent
satisce in choosing an action.
So, following similar ideas by Behn and Vaupel (1982), Vaupel (1986), and
others, in my view a better normative theory of bounded rationality recognizes
that the agent should satisce in the construction of a bounded model. But
within that model the theory recommends the agent to choose the best possible
action. This seems quite dierent from Simons recommendation that the agent
satisce in choosing an action.

19 Stochastic Descriptive Models

So far, I have been almost exclusively concerned with normative standards of
rational behaviour. But as the discussion of bounded rationality makes clear,
the SEU model pays little or no attention to psychological reality. Of course,
psychological reality seems more relevant for describing actual behaviour which
is generally not rational than it is for thinking about normative standards of
rational behaviour. Nevertheless, normative standards that pay no attention at
all to psychological reality can easily be dismissed as irrelevant. This is especially
so in game theory, where one can imagine oering advice to one player on how
to behave rationally when some or all of the other players are not necessarily
rational. Such advice should be based on a model which describes realistically
the behaviour of all but the one player who is receiving the advice.
Constructing a realistic descriptive model of behaviour is perhaps more of
a task for psychologists than for economists. And I know even less about psy-

30
chology than I do about philosophy. Nevertheless, it seems that only stochastic
models oer any serious chance of describing behaviour at all well. Moreover,
the following slight variation of a well known example due to McFadden (1974)
illustrates how, in any realistic stochastic model of behaviour, the structure of
the decision tree is very likely to be important.
Suppose somebody is contemplating going by public transport from Milan
to, say, Como. Suppose too that this person can go either to the railway station
and catch a train, or else to the bus station and catch one of two buses red or
blue. Now, if the railway and bus stations are very close, the traveller may well
have time to examine the timetables for the trains and for both red and blue
buses before making a single decision determining how to travel to Como. In
this case, lacking further information, we might think that there is a 1/3 chance
of the agent taking the train, or either kind of bus, depending on when the
agent reached the station, the respective fares, which train or bus is scheduled
to reach Como rst, etc.
Suppose on the other hand that the railway and bus stations are suciently
far apart that the traveller will certainly depart from whichever station he goes
to rst. Then the decision is made in two stages. First the traveller decides
whether to head for the railway or bus station. At the railway station there is
no further decision to make, but at the bus station there is the second decision
between the red and blue buses. In this case the probability of the traveller
going to the railway station and taking a train may well rise to 1/2, whereas
with probability 1/2 he goes over to the bus station. If the probabilities of
catching either kind of bus remain equal, both drop to 1/4. For more details,
see Amemiya (1981, 1985).
Such behaviour has consequences which depend on the structure of the de-
cision tree. So the consequentialist axiom is violated. The example brings out
important dierences between, on the one hand, normative models of rational
behaviour and, on the other hand, descriptively accurate models that may not
be rational at all.

20 Confusing Normative and Descriptive

Most theoretical work in economics, even when it is concerned with description
or prediction rather than prescription, is actually based on the standard norma-
tive model. That is, it postulates preference maximizing consumers, and prot
maximizing rms. Often consumers are assumed to maximize expected utility
and rms to maximize expected prot, though there has recently been extensive
interest in models where consumers maximize non-expected utility.
This confusion of normative and descriptive seems to me a fundamental
methodological error. I shall not try to defend it, though I will venture four
positive explanations. Of these, the rst is rather obvious: inertia or laziness.
Demand theory in economics has always been based on utility maximization,
though utility was rst rened to ordinal utility and then virtually dispensed
with through the revealed preference approach. Nevertheless, the neoclassical

31
utility maximization approach to the theory of consumer demand gets taught
to succeeding generations of students, and so passed on.
A second possible explanation is mathematical sophistication and clarity.
The remarkable intellectual revolution in economics which occurred during the
1950s owed much to the use of mathematical techniques. It will be hard for
anything as simple and messy as most stochastic choice models to supplant the
familiar maximization models of modern neoclassical theory.
A third possible explanation is one I owe to Joan Robinson, and her book
Economic Philosophy. It is the suggestion that many economists have been
little more than apologists for the laisser faire free market system. This is per-
haps closer to the truth in the U.S.A., where many economists teach at business
schools, and even more receive direct or indirect nancial support from business
people. Other economists outside business schools may nevertheless consult,
teach people who are planning business careers, teach at universities that rely
on business prots to endow their chairs, or simply teach at universities where
students have their high tuition fees paid by parents whose incomes are derived
from business. Given all these nancial pressures, it is perhaps somewhat sur-
prising that there are nevertheless many critics of the free enterprise system
who ourish in U.S. universities.
Anyway, whatever their motivation, whenever defenders of free enterprise
care to think at all about the performance of the economic system as a whole,
they love to invoke the fundamental eciency theorems of welfare economics.
Under important qualications which are often conveniently forgotten, these
theorems demonstrate that free markets are associated with Pareto ecient
outcomes in which it is impossible to make any one consumer better o without
making some others worse o. Yet even when they are valid, these eciency
theorems are totally devoid of normative content unless one postulates that
consumers choose what is right for themselves, by rationally maximizing an
appropriate preference ordering (or ordinal utility function). Another favourite
result concerns the (potential) gains from free trade, enhanced competition,
privatization, or other forms of economic liberalization. Yet this result also
relies on the same kind of maximization postulate. Defenders of free enterprise
would have the force of their arguments considerably reduced if consumers were
known to be behaving irrationally. Yet then economists might have a new role
to play advising consumer organizations instead of business. Perhaps, however,
they see that business can aord to pay better.
A fourth reason for using the model of rational behaviour descriptively could
be that it has been tested experimentally. This would be the best reason, of
course, if it were valid. However, the experimental evidence to date seems
to be rather mixed. Where experiments face subjects with relatively simple
decision problems with clear monetary objectives, and also give subjects an
opportunity to learn from experience, there is some evidence that behaviour
converges to what would be predicted in the (often unique) equilibrium of a
game theoretic model. This seems to be especially true of the experimental
markets studied by Plott and others see Sunders (1995) survey, for instance.
Outside some rather special laboratory situations, however, there seems to be

32
little evidence of rationality in actual behavior. This is true even for special
classes of experimental subjects like economics or business students, who are
supposed to understand something of what it means to be rational.

21 Problems Solved and Unsolved

I feel that my own understanding of what constitutes rationality in economics
has progressed far since I rst became interested in the topic about 30 years ago.
Then, I remember being somewhat horried by the rst economics textbook I
read, which happened to be Samuelsons Economics, because of the way in which
consumer behaviour was reduced to analysing an indierence map involving only
that one consumers personal consumption.
The theory of rational behavior itself had made huge conceptual advances
in the period from the mid 1940s, when von Neumann and Morgenstern pub-
lished their book, until 1963, when Anscombe and Aumanns important work
appeared. However, the standard axioms were often heavily criticized, even as
a basis for a normative theory. So the challenge had become how to investi-
gate the foundations more carefully, by considering what might lie behind the
standard axioms.
With this in mind, the three consequentialist axioms and their ancillaries,
applied to behaviour in single person decision trees, do appear to oer a more
secure foundation for standard normative decision theory. This is especially true
of the SEU model in the form proposed by Anscombe and Aumann (1963). But
there are problems especially with the unrestricted domain assumption when
the agent may have reason to favour some decision trees over others because
they allow choices to be deferred and so oer more exibility. Nevertheless,
this consequentialist approach has been attered by the attention of several
good humoured and sympathetic critics such as Machina (1989), McClennen
(1990) and Munier (1996). It has been an interesting challenge of late, because
of the signicant developments in non-expected utility models of choice under
risk and in alternatives to subjective probability. But on the whole the non-
standard models purport to be descriptive, whereas I claim that the SEU model
is suitable as a normative standard only. Saying that often enough takes much
of the sting out of the criticisms.
For me at least, the exploration of normative single-person decision theory
seems almost over. There may be interesting technical problems involved in
extending the theory to innite decision trees, and/or those with an innite
horizon. But these are mathematical niceties. More pressing issues arise in
dealing with games. Even for normal form games, the work on justifying the
SEU* approach remains too novel for me to be condent yet of its ultimate
acceptability. But for games in extensive form there are more serious issues. For
instance, there is recent unpublished work by Asheim and Dufwenberg (1996)
claiming that players may have more than one set of rationalizable strategies.
One should also reconcile the need for positive probabilities with the desire to

33
attach zero probabilities to strategies that are not rationalizable. Serious open
questions like these remain incompletely resolved.
Last but not least is the urgent need to develop better concepts of ratio-
nality than the structural consistency conditions which currently dominate the
economics literature. There is certainly a need for alternative normative models
that allow us to say more about which beliefs and preferences are rational. And
for normative models which make fewer psychological demands on agents, as
well as recognizing psychological reality better than current models seem to.
Although I do maintain that the fundamental separation between normative
and descriptive models of behaviour remains necessary.

Acknowledgements
November 1997 nal version of material prepared for a lecture presented at
the Universit`
a Cattolica del Sacro Cuore, Milan, on 28th March 1996. Several
parts are based on draft chapters to appear in Salvador Barber`
a, Peter J. Ham-
mond and Christian Seidl (eds.) Handbook of Utility Theory (in preparation
for Kluwer Academic Publishers). Section 15 in particular has beneted from
the comments of Marco Mariotti on a draft chapter for the Handbook.

References
Afriat, S.N. (1967) The Construction of a Utility Function from Demand
Data International Economic Review, 8, 6777.
Akerlof, G.A. and J.L. Yellen (1985a) Can Small Deviations from Rationality
Make Signicant Dierences to Economic Behavior? American Economic
Review, 75, 708720.
Akerlof, G.A. and J.L. Yellen (1985a) A Near-Rational Model of the Business
Cycle, with Wage and Price Inertia Quarterly Journal of Economics, 100,
823838.
Allais, M. (1953) Le comportement de lhomme rationnel devant le risque:
critique des postulats et des axiomes de lecole americaine, Econometrica,
21, 503546; summary version of paper whose translation appeared as Allais
(1979a).
Allais, M. (1979a) The Foundations of a Positive Theory of Choice Involving
Risk and a Criticism of the Postulates and Axioms of the American School,
in Allais and Hagen (1979), pp. 27145.
Allais, M. (1979b) The So-called Allais Paradox and Rational Decisions under
Uncertainty, in Allais and Hagen (1979), pp. 437681.
Allais, M. (1987) Allais Paradox, in Eatwell, J., Milgate, M. and P. Newman
(eds.) The New Palgrave: A Dictionary of Economics (London: Macmillan,
1987); reprinted in Eatwell, J., Milgate, M. and P. Newman (eds.) The New
Palgrave: Utility and Probability (London: Macmillan, 1990), pp. 39.
Allais, M. and O. Hagen (eds.) (1979) The Expected Utility Hypothesis and
the Allais Paradox (Dordrecht: D. Reidel).

34
Amemiya, T. (1981) Qualitative Response Models: A Survey Journal of
Economic Literature, 19, 14831536.
Amemiya, T. (1985) Advanced Econometrics (Oxford: Basil Blackwell).
Anscombe, F.J. and R.J. Aumann (1963) A Denition of Subjective Probabil-
ity Annals of Mathematical Statistics, 34, 199205.
Anscombe, G.E.M. (1958) Modern Moral Philosophy Philosophy, 33, 119.
Armbruster, W. and W. B oge (1979) Bayesian Game Theory, in O. Moeschlin
and D. Pallaschke (eds.) Game Theory and Related Topics (Amsterdam:
North-Holland), pp. 1728.
Arrow, K.J. (1951) Alternative Approaches to the Theory of Decision in
Risk-Taking Situations, Econometrica, 19, 404437; reprinted in Arrow
(1971), ch. 1 and in Arrow (1984), ch. 2.
Arrow, K.J. (1959) Rational Choice Functions and Orderings Economica, 26,
121127; reprinted in Arrow (1984), ch. 6.
Arrow, K.J. (1963) Social Choice and Individual Values (2nd. edn.) (New
Haven: Yale University Press).
Arrow, K.J. (1971) Essays in the Theory of Risk-Bearing (Chicago: Markham;
Amsterdam: North-Holland).
Arrow, K.J. (1984) Collected Papers of Kenneth J. Arrow, Vol. 3: Individual
Choice under Certainty and Uncertainty (Cambridge, Mass.: Belknap Press
of Harvard University Press).
Arrow, K.J., E. Colombatto, M. Perlman, and C. Schmidt (eds.) (1996) The
Rational Foundations of Economic Behaviour (London: Macmillan).
Asheim, G. and M. Dufwenberg (1996) Admissibility and Common Knowl-
edge, preprint, Department of Economics, University of Oslo.
Aumann, R.J. (1987) Correlated Equilibrium as an Expression of Bayesian
Rationality, Econometrica, 55, 118.
Bandyopadhyay, T. (1978) Revealed Preference Theory, Ordering and the
Axiom of Sequential Path Independence Review of Economic Studies, 55,
343351.
Battigalli, P. (1996) Comment, on Mariotti (1996), in Arrow et al. (eds.)
(1996), pp. 149154.
Behn, R.D. and J. W. Vaupel (1982) Quick Analysis for Busy Decision Makers
(New York: Basic Books).
Bernheim, B.D. (1984) Rationalizable Strategic Behavior Econometrica, 52,
10071028.
Bernheim, B.D. (1986) Axiomatic Characterizations of Rational Choice in
Strategic Environments, Scandinavian Journal of Economics, 88, 473488.
Bernoulli, D. (1738) Specimen theoriae novae de mensura sortis in Com-
mentarii Academiae Scientiarum Imperialis Petropolitanae; translated by
Sommer (1954).

35
Binmore, K. (1987) Modelling Rational Players, I, Economics and Philosophy,
3, 179214; reprinted (1990) in Essays on the Foundations of Game Theory
(Cambridge, Mass.: Basil Blackwell), ch. 5, pp. 151185.
Blume, L., A. Brandenburger and E. Dekel (1991a) Lexicographic Probabilities
and Choice Under Uncertainty Econometrica, 59, 6179.
Blume, L., A. Brandenburger and E. Dekel (1991b) Lexicographic Probabilities
and Equilibrium Renements Econometrica, 59, 8198.
Boge, W. and T. Eisele (1979) On Solutions of Bayesian Games, International
Journal of Game Theory, 8, 193215.
Borel, E. (1924) Sur les jeux o`
u interviennent lhasard et lhabilite des joueurs
in J. Hermann (ed.) Elements de la theorie des probabilites (Paris: Librairie
Scientique), 204224.
Borgers, T. (1993) Pure Strategy Dominance, Econometrica, 61, 423430.
Campbell, D.E. (1978) Realization of Choice Functions Econometrica, 46,
171180.
Cramer, G. (1728) Letter to Nicholas Bernoulli; extracts printed in translation
by Sommer (1954).
Davidson, D., J.C.C. McKinsey and P. Suppes (1955) Outlines of a Formal
Theory of Value, I Philosophy of Science, 22, 140160.
De Finetti, B. (1937) La prevision: ses lois logiques, ses sources subjectives,
Annales de lInstitut Henri Poincare, 7; translated as Foresight: Its Logical
Laws, Its Subjective Sources, in Kyburg and Smokler (1964), pp. 93158.
Elster, J. (1979) Ulysses and the Sirens (Cambridge: Cambridge University
Press).
Foley, D.K. (1970) Lindahls Solution and the Core of an Economy with Public
Goods Econometrica, 38, 6672.
Gul, F. (1996) Rationality and Coherent Theories of Strategic Behavior
Journal of Economic Theory, 70, 131.
Hammond, P.J. (1976) Changing Tastes and Coherent Dynamic Choice,
Review of Economic Studies, 43, 159173.
Hammond, P.J. (1977) Dynamic Restrictions on Metastatic Choice, Econom-
ica, 44, 337350.
Hammond, P.J. (1988a) Consequentialist Foundations for Expected Utility,
Theory and Decision, 25, 2578.
Hammond, P.J. (1988b) Orderly Decision Theory: A Comment on Professor
Seidenfeld, Economics and Philosophy, 4, 292297.
Hammond, P.J. (1991) Interpersonal Comparisons of Utility: Why and
How They Are and Should Be Made, in J. Elster and J.E. Roemer
(eds.), Interpersonal Comparisons of Well-Being (Cambridge: Cambridge
University Press), ch. 7, pp. 200254.

36
Hammond, P.J. (1993) Aspects of Rationalizable Behavior, in Binmore, K.,
Kirman, A.P. and P. Tani (eds.) Frontiers of Game Theory (Cambridge,
Mass.: M.I.T Press), ch. 14, pp. 277305.
Hammond, P.J. (1994) Elementary Non-Archimedean Representations of
Probability for Decision Theory and Games, in Humphreys, P. (ed.)
Patrick Suppes: Scientic Philosopher, Vol. I: Probability and Probabilistic
Causality (Dordrecht: Kluwer Academic Publishers), ch. 2, pp. 2559.
Hammond, P.J. (1996) Consequentialist Decision Theory and Utilitarian
Ethics, in F. Farina, F. Hahn, and S. Vannucci (eds.) Ethics, Rationality,
and Economic Behaviour (Oxford: Clarendon Press, 1996), pp. 92118.
Hammond, P.J. (1997) Consequentialism, Non-Archimedean Probabilities,
and Lexicographic Expected Utility, to appear in C. Bicchieri, R. Jerey
and B. Skyrms (eds.) The Logic of Strategy (Oxford University Press).
Harsanyi, J.C. (1955) Cardinal Welfare, Individualistic Ethics, and Interper-
sonal Comparisons of Utility, Journal of Political Economy, 63, 309321.
Harsanyi, J.C. (19678) Games with Incomplete Information Played by
Bayesian Players, IIII, Management Science, 14, pp. 159182, 320334,
486502.
Hicks, J.R. and R.G.D. Allen (1934) A Reconsideration of the Theory of
Value, I and II Economica, 1, 5276 and 196219.
Hillas, J. (1990) On the Denition of the Strategic Stability of Equilibria
Econometrica, 58, 13651390.
Houthakker, H.S. (1950) Revealed Preference and the Utility Function
Economica, 17, 159174.
Howard, J. (1988) Cooperation in the Prisoners Dilemma Theory and
Decision, 24, 203213.
Keynes, J.M. (1921) A Treatise on Probability (London: Macmillan).
Kohlberg, E. and J.-F. Mertens (1986) On the Strategic Stability of Equilibria,
Econometrica, 54, 10031037.
Kreps, D., Milgrom, P., Roberts, J. and R. Wilson (1982) Rational Cooperation
in the Finitely Repeated Prisoners Dilemma, Journal of Economic Theory,
27, 245252.
Kreps, D. and R. Wilson (1982) Sequential Equilibrium, Econometrica, 50,
863894.
Kyburg, H.E. and H.E. Smokler (1964) Studies in Subjective Probability (New
York: John Wiley).
Luce, R.D. and H. Raia (1957) Games and Decisions: Introduction and Critical
Survey (New York: John Wiley).
Machina, M.J. (1989) Dynamic Consistency and Non-Expected Utility Models
of Choice Under Uncertainty, Journal of Economic Literature, 28, 1622
1668.

37
Malinvaud, E. (1972), Lectures on Microeconomic Theory (Amsterdam: North
Holland).
Mariotti, M. (1996) The Decision Theoretic Foundations of Game Theory, in
Arrow et al. (eds.) (1996), ch. 6, pp. 133148.
McClennen, E.F. (1990) Rationality and Dynamic Choice: Foundational
Explorations (Cambridge: Cambridge University Press).
McFadden, D. (1974) Conditional Logit Analysis of Qualitative Choice
Behavior, in P. Zarembka (ed.) Frontiers in Econometrics (New York:
Academic Press), pp. 105142.
McLennan, A. (1989) Consistent Conditional Systems in Noncooperative
Game Theory International Journal of Game Theory, 18, 141174.
Milleron, J.-C. (1972) Theory of Value with Public Goods: A Survey Article
Journal of Economic Theory, 5, 419477.
Mongin, P. (1994) Loptimisation est-elle un crit`ere de rationalite individuelle?
Dialogue, 33, 191222.
Morgenstern, O. (1935) Vollkommene Voraussicht und wirtschaftliches Gle-
ichgewicht Zeitschrift f
ur National
okonomie, 6, pp. 337357; translated by
F.H. Knight as Perfect Foresight and Economic Equilibrium in A. Schot-
ter (ed.) Selected Writings of Oskar Morgenstern (New York: New York
University Press, 1976), ch. 10, pp. 169183.
Munier, B. (1996) Comment, in Arrow et al. (eds.) (1996), pp. 4347.
Myerson, R.B. (1978) Renements of the Nash Equilibrium Concept,
International Journal of Game Theory, 7, 7380.
Nash, J.F. (1950) Equilibrium Points in n-Person Games, Proceedings of the
National Academy of Sciences of the USA, 36, 4849.
Nash, J.F. (1951) Non-Cooperative Games, Annals of Mathematics, 54,
286295.
Nau, R.F. and K.F. McCardle (1990) Coherent Behavior in Noncooperative
Games Journal of Economic Theory, 50, 424444.
Osborne, M.J. (1990) Signaling, Forward Induction, and Stability in Finite
Repeated Games Journal of Economic Theory, 50, 2236.
Osborne, M.J. and A. Rubinstein (1994) A Course in Game Theory (Cambridge,
Mass.: MIT Press).
Pearce, D. (1984) Rationalizable Strategic Behavior and the Problem of
Perfection, Econometrica, 52, 10291050.
Piccione, M. and A. Rubinstein (1994) On the Interpretation of Decision
Problems with Imperfect Recall University of Tel Aviv, Sackler Institute
for Economic Studies, Working Paper No. 24/94.
Plott, C.R. (1973) Path Independence, Rationality and Social Choice
Econometrica, 41, 10751091.

38
Ramsey, F.P. (1926) Truth and Probability, in The Foundations of Mathe-
matics and Other Logical Essays (London: Kegan Paul, 1931), pp. 156198;
reprinted in Kyburg and Smokler (1964) and in G ardenfors, P. and N.-E.
Sahlin (eds.) (1988) Decision, Probability, and Utility (Cambridge: Cam-
bridge University Press).
Rosenthal, R.W. (1981) Games of Perfect Information, Predatory Pricing and
the Chain-Store Paradox, Journal of Economic Theory, 25, 92100.
Samuelson, P.A. (1938) A Note on the Pure Theory of Consumers Behavior
and : An Addendum Economica, 5, 6171 and 3534.
Savage, L.J. (1954) The Foundations of Statistics (New York: John Wiley).
Schelling, T.C. (1960) The Strategy of Conict (Cambridge, MA: Harvard
University Press).
Selten, R. (1965) Spieltheoretische Behandlung eines Oligopolmodells mit
Nachfragetragheit Zeitschrift f
ur die gesamte Staatswissenschaft, 121,
301324 and 667689.
Selten, R. (1978) The Chain Store Paradox, Theory and Decision, 9, 127159.
Simon, H.A. (1955) A Behavioral Model of Rational Choice Quarterly Journal
of Economics, 69, 99118.
Simon, H.A. (1957) Models of Man (New York: John Wiley).
Sommer, L. (1954) Exposition of a New Theory on the Measurement of Risk,
(translation) Econometrica, 22, 2336.
Sunder, S. (1995) Experimental Asset Markets: A Survey in J.H. Kagel
and A.E. Roth (eds.) Handbook of Experimental Economics (Princeton:
Princeton University Press), ch. 6, pp. 445500.
Tan, T.C.-C. and S.R. da C. Werlang (1988) The Bayesian Foundations of
Solution Concepts of Games Journal of Economic Theory, 45, 370391.
Uzawa, H. (1960) Preference and Rational Choice in the Theory of Consump-
tion in K.J. Arrow, S. Karlin and P. Suppes (eds.) Mathematical Methods
in the Social Sciences, 1959, (Stanford: Stanford University Press), ch. 9.
van Damme, E. (1989) Stable Equilibria and Forward Induction Journal of
Economic Theory, 48, 476496.
Vaupel, J.W. (1986) Un pensiero analitico per decisori indaarati in L. Sacconi
(ed.) La decisione: Razionalit`a collettiva e strategie nellamministrazione e
nelle organizzazioni (Milano: Franco Angeli, 1986), ch. 13, pp. 226238.
von Neumann, J. (1928) Zur Theorie der Gesellschaftsspiele, Mathematische
Annalen, 100, 295320; reprinted in A.H. Taub (ed.) Collected Works of
John von Neumann, Vol. VI (Oxford: Pergamon Press, 1963), pp. 126.
von Neumann, J. and O. Morgenstern (1944; 3rd edn. 1953) Theory of Games
and Economic Behavior (Princeton: Princeton University Press).

Zermelo, E. (1912) Uber eine Anwendung der Mengenlehre auf die Theorie
der Schachspiels Proceedings of the Fifth International Congress of
Mathematicians, Vol. II, 501504.

39